Purchase this article with an account.
Jeroen Joukes, Bart Krekelberg; A motion detection model based on a recurrent network. Journal of Vision 2011;11(11):746. doi: 10.1167/11.11.746.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Current motion models such as the motion energy (ME) model rely on specific, fixed neural delays between cells in the motion pathway. It is not clear whether neurons in the motion processing pathway of primates have the temporal specificity to support this assumption. Our goal was to investigate the feasibility of motion detection using a recurrent neural network without carefully tuned delays. We recorded the responses of neurons in the middle temporal (MT) area of macaques to patterns moving in the preferred or anti-preferred direction at different speeds. Next, we trained a recurrent neural network to reproduce the responses to motion in the preferred direction. The network accurately captured the temporal dynamics and speed tuning. This shows that explicit temporal delays are not needed to reproduce typical motion responses; they can be implemented with recurrent connections. We further tested the network with methods commonly used to analyze real neural data. We found that the network generalized to all moving input patterns (pattern invariance). Moreover, even though the network was trained only on motion in the preferred direction, it generalized to the anti-preferred response. Third, spike triggered covariance revealed filters similar to those in the ME model. Two excitatory filters in anti-phase had a space time slant that matched the preferred speed and direction of the output. Two inhibitory filters in anti-phase had a slant that matched the low speed anti-preferred direction. Taken together these results show that a recurrent neural network can reproduce the tuning of motion sensitive cells. When probed with standard methods, this network behaves much like the ME model, even though none of the ME stages can be mapped directly onto this architecture. In other words, the computation of the recurrent network is equivalent, but the underlying hardware is fundamentally different and, we believe, more biologically plausible.
This PDF is available to Subscribers Only